In This Story
Online shopping is exploding, with more customers double-clicking instead of wandering store aisles. The drawback to online shopping is the inability to touch items, to feel the fabric or inspect the shoes. Do they look cheap? Do they run small? Is that vacation destination nice, or are the rooms stuffy? Online businesses rely on customer reviews to offer these answers.
User reviews often comprise two parts, the starred rating and the review. Jingyuan Yang, assistant professor of information systems and operations management, noticed a problem in that system. In her paper, “NeuO: Exploiting the Sentimental Bias between Ratings and Reviews with Neural Networks” (with coauthors Yuanbo Xu, Yongjian Yang, Jiayu Han, En Wang, Fuzhen Zhuang, and Hui Xiong), she notes that often the review is missing or doesn’t match up with the rating. This gap is problematic because, she says, “It is really important that users’ ratings and reviews be mutually reinforced to grasp the users’ true opinions.”
Yang and her coauthors exploited two-step training neural networks, using both reviews and ratings to grasp users’ true opinions. They developed an opinion-mining model using a specialized linear mathematical operation called convolution to ensure ratings. They used a combination function designed to catch the opinion bias and proposed a recommendation method using the enhanced user-item matrix.
Virtual businesses need healthy user reviews. When customers don’t feel they can rely on reviews, their trust in the company falters. Yang and her team have helped shore up the review system, an effort that will go a long way toward building happy customer bases.